AI-Assisted Papillary Thyroid Carcinoma Localization on Cervical Non-contrast CT: A Multicenter Validation and Two-Reader Pilot Study of a Standalone Executable nnU-Net Software.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The Fourth Affiliated Hospital of Nanjing Medical University, Nanjing, 210031, People's Republic of China.
- Department of Radiology, Zhongda Hospital Southeast University, Nanjing, 210048, People's Republic of China.
- Department of Central Laboratory, The Fourth Affiliated Hospital of Nanjing Medical University, 210031, Nanjing, People's Republic of China.
- Department of Nuclear Medicine, Nanjing Drum Tower Hospital, The Affiliated Hospital of Medicine school, Nanjing University, 210008, Nanjing, People's Republic of China. [email protected].
- Department of General Surgery, The Fourth Affiliated Hospital of Nanjing Medical University, 210031, Nanjing, People's Republic of China. [email protected].
Abstract
Accurate preoperative localization and 3D anatomical assessment of papillary thyroid carcinoma (PTC) are critical for surgical decision-making. Cervical non-contrast CT (NCCT) avoids iodinated contrast-related risks but has limited PTC lesion conspicuity, while first-line ultrasound suffers from operator dependence and incomplete 3D anatomical information. We developed a 3D full-resolution (fullres) nnU-Net model for joint thyroid and PTC segmentation on cervical NCCT, and conducted a pilot reader study to evaluate its auxiliary value for a junior and a senior radiologist. This multicenter retrospective study included 300 cases for model training, 120 temporal and 100 external test cases. Model performance was assessed via standard segmentation metrics, and the optimal model was packaged as a standalone executable Windows software. The 3D fullres model significantly outperformed 2D and single-label models for both PTC and thyroid segmentation (all P < 0.001). AI assistance significantly improved the diagnostic sensitivity of the participating junior radiologist (P < 0.001) and reduced active human reading time by > 40% for both readers, while the human-in-the-loop workflow effectively mitigated the model's inherent false-positive risk. This first NCCT-based 3D nnU-Net joint segmentation model improves the PTC diagnostic efficiency and sensitivity of the two participating radiologists, with an open-source standalone executable software providing a novel contrast-free tool to assist in regional PTC localization prior to surgery. Trial Registration: retrospectively registered in the Chinese Clinical Trial Registry (ChiCTR2500097848) on 2025-02-26.